Inter-hemispheric integration of visual processing during task-driven lateralization

Abstract

The mechanisms underlying inter-hemispheric integration (IHI) remain poorly understood, particularly for lateralized cognitive processes. To test competing theories of IHI, we constructed and fitted dynamic causal models to functional magnetic resonance data from two visual tasks that operated on identical stimuli but showed opposite hemispheric dominance. Using a systematic Bayesian model selection procedure, we found that in the ventral visual stream, which was activated by letter judgments, inter-hemispheric connections mediated asymmetric information transfer from the non-specialized right to the specialized left hemisphere when the latter did not have direct access to stimulus information. Notably, this form of IHI did not engage all areas activated by the task but was specific for areas in the lingual and fusiform gyri. In the dorsal stream, activated by spatial judgments, it did not matter which hemisphere received the stimulus: Inter-hemispheric coupling increased bidirectionally, reflecting recruitment of the non-specialized left hemisphere. Again, not all areas activated by the task were involved in this form of IHI; instead, it was restricted to interactions between areas in the superior parietal gyrus. Overall, our results provide direct neurophysiological evidence, in terms of effective connectivity, for the existence of context-dependent mechanisms of IHI that are implemented by specific visual areas during task-driven lateralization.

Introduction

The nature of hemispheric specialization is an old enigma. One particularly vexing question is how the brain integrates processes that are lateralized to opposite hemispheres (Johansson et al. 2006). Studies of patients with callosal lesions (Gazzaniga 2000) and of healthy volunteers (Hellige 1990) imply that even closely related tasks can rely on different mechanisms of inter-hemispheric integration (IHI). For example, controlling the focus of spatial attention and identifying spatial locations are joint sub-processes of many visuospatial tasks, yet they show strikingly different dependencies on callosal integrity (Holtzman et al. 1981). This complexity calls for computational models of IHI that account for context- and task-dependencies (Hellige 1990; Liederman 1998). The challenge is to test such models, directly and quantitatively, for the human brain: Given a particular lateralized task, which candidate mechanism of IHI best explains the measured neural responses?

A powerful characterization of integration mechanisms in neural systems is through “effective connectivity”, i.e. the causal influences that system elements exert over one another (Friston 1994; Horwitz et al. 1999). Specifically, to better understand IHI, one needs anatomically precise models that specify how individual connections change with cognitive context (task requirements and/or stimulus properties). Given the reciprocal nature of callosal connections and the ensuing complexity of inter-hemispheric models, this approach was previously hampered by methodological limitations (but see McIntosh et al. 1994).

Recently, dynamic causal modeling (DCM) was introduced as a new approach to inferring effective connectivity from fMRI data (Friston et al. 2003). In a single-subject analysis, we previously demonstrated the usefulness of DCM for complex inter-hemispheric models (Stephan et al. 2005). Here, we report an extended analysis of a group of subjects (Fig. 1), combining DCM with Bayesian model selection (BMS), to re-analyze fMRI data from a novel paradigm that probes IHI during two inversely lateralized tasks, a letter decision (LD) and a spatial decision (SD) task, using identical and peripherally presented visual stimuli (see Fig. 1). An fMRI study by Stephan et al. (2003) found that LD activated the left ventral stream of the visual system whereas SD activated the right dorsal stream (Ungerleider & Mishkin 1982; Merigan & Maunsell 1993).

Figure 1.

Summary of the experimental design and SPM results. See Methods and Supplemental Material for details. The right part of this figure shows the results of an analysis with SPM2 (p<0.05 whole-brain cluster-level corrected): Despite identical stimuli, strong task-dependent lateralization of activity was found. In particular, letter decisions activated the left ventral stream of the visual system whereas spatial decisions activated the right dorsal stream (ellipsoids).

In this paradigm, IHI could be explained by three well-established theories of inter-hemispheric interactions: asymmetric information transfer, inter-hemispheric inhibition, and hemispheric recruitment. These theories predict different patterns of inter-hemispheric effective connectivity as a neurophysiological signature of IHI (see Discussion for details). We investigated which theory best described IHI (i) in the ventral stream of the visual system during letter judgments and (ii) in the dorsal stream during spatial judgments, and (iii) which visual areas were at all involved in IHI during our tasks. 64 candidate DCMs were fitted for each subject; from these, BMS determined optimal models of IHI in ventral and dorsal streams (Fig. 1). Our results suggest that, for the particular paradigm studied, IHI is characterized by asymmetric information transfer between lingual and fusiform gyri in the ventral stream and by hemispheric recruitment in the dorsal stream, involving the superior parietal gyrus.

Materials and Methods

Experimental design and data acquisition

Fig. 1 summarizes the most important aspects of the experimental design and stimulus presentation. The central idea of this paradigm is to apply two tasks requiring language and visuospatial processes, respectively, to peripherally presented stimuli that contain both language and spatial features. The stimuli were concrete, high-frequency German nouns, each consisting of four letters. In each stimulus either the second or third letter was red whereas all other letters were black. During a letter decision (LD) task, the participants indicated by button press whether the word contained the target letter “A” or not. During a spatial decision (SD) task, they indicated by button press whether the single red letter of the word was located left or right from the midline of the word. 16 right-handed male adult subjects were scanned on a 1.5T Siemens Sonata scanner. All details of experimental design and data acquisition can be found in Stephan et al. (2003) and in the Supplemental Material.

Construction of statistical parametric maps

The results previously reported by Stephan et al. (2003) were obtained using the software package SPM99. To enable the application of DCM and BMS, we re-analyzed the data using SPM2 and adopted a slightly different preprocessing strategy. For this reason, the activation maps shown in Fig. 1 are slightly different from those published in Stephan et al. (2003). Briefly summarized, the present analysis of the fMRI data comprised the following steps. For each subject, after discarding the first five images, the remaining 700 images were realigned to correct for head movements, spatially normalized to the Montreal Neurological Institute (MNI) template brain, smoothed spatially with a three-dimensional Gaussian kernel of 8 mm full width half maximum (FWHM) and resampled, resulting in 3×3×3 mm voxels. The data were then modeled voxel-wise, using a general linear model that included all combinations of task, visual field, and response hand, plus effects of no interest (instruction periods and realignment parameters to account for motion-related variance). The data were high-pass filtered (cut-off 1/128 s) to remove low-frequency signal drifts. A first-order autoregressive model was used to remove serial correlation in the data. Contrast images were created for each subject and entered separately into voxel-wise one-sample t-tests (df = 15), implementing a random effects analysis. The statistical threshold was set at p<0.05 at the cluster-level (with a standard voxel-level cut-off of p<0.001), whole-brain corrected for family-wise errors (FWE) using Gaussian Random Field Theory (Poline et al. 1997).

Dynamic causal modeling (DCM)

DCM is an established model of neural system dynamics that has been methodologically evaluated by several studies (Friston et al. 2003; Lee et al. 2006; Penny et al. 2004a, 2004b) and has found numerous applications (e.g. Bitan et al. 2005; Haynes et al. 2005; Mechelli et al. 2003; Smith et al. 2006; Stephan et al. 2005). Given some measured regional fMRI time series, DCM enables one to infer the connectivity between the neural sources that give rise to these regional measurements. The basic idea is to estimate the parameters of a reasonably realistic neural model such that the predicted regional BOLD signals, which result from converting the modeled neural dynamics into hemodynamic responses, correspond as closely as possible to the observed BOLD signals. Importantly, DCM models how the neural dynamics are shaped by experimentally controlled manipulations, i.e. external inputs u, that enter the model in two different ways. Inputs can elicit responses through direct influences on specific regions (“driving inputs”, e.g. sensory inputs) or they can change the strength of coupling among regions (“modulatory inputs”, e.g. task effects or learning). This distinction represents an analogy, at the level of neural populations, to the concept of driving and modulatory afferents in studies of single neurons (Sherman & Guillery 1998).

Mathematically, DCM is based on a bilinear model of neural population dynamics that is combined with a hemodynamic model describing the transformation of neural activity into predicted BOLD responses (Buxton et al. 1998; Friston et al. 2000). The “hidden” neural dynamics (i.e. not directly observed by fMRI) are modeled by the following bilinear differential equation:

$$\frac{\mathit{dz}}{\mathit{dt}}=\left(A+\underset{j=1}{\overset{m}{\Sigma }}{u}_j{B}^{\left(j\right)}\right)z+\mathit{Cu}$$ (1)

Here, z is the state vector (with each state variable representing the population activity of one region in the model), t is continuous time, and u_j is the j-th input to the modelled system (i.e. some experimentally controlled manipulation). In this state equation, the A matrix contains the “intrinsic” or “fixed” connection strengths between the modeled regions, and the B⁽¹⁾…B^(m) matrices represent the context-dependent modulation of these connections, e.g. by task, as an additive change. Finally, the C matrix represents the strengths of direct (“driving”) inputs to the modeled system (e.g. sensory stimuli). Note that all parameters correspond to rate constants of the modeled neurophysiological processes and are thus in units of 1/s = Hz.

For any given set of parameter values, the neural state equation can be integrated, and the resulting neural dynamics transformed into predicted BOLD signals, using a well-established hemodynamic model (Friston et al. 2000). Combining the neural and hemodynamic state equations into a joint forward model, DCM uses Bayesian inversion to determine the posterior densities of the parameters (Friston et al. 2003). Under Gaussian assumptions (Laplace approximation), these densities can be characterized in terms of their maximum a posteriori (MAP) estimates and their posterior covariances.

Practical implementation of DCM in this study

(i) Choice of areas and time series extraction

The definition of areas was informed by the results of the conventional SPM analysis; see the Results section for details. DCMs are fitted to subject-specific BOLD time series. Because the exact locations of activated areas vary over subjects, a general challenge is to define the elements of the modeled system (and thus the extracted time series) such that models are comparable across subjects. Here, we ensured comparability across subjects by requiring that the extracted time series met a combination of anatomical and functional criteria (see Supplemental Material for details). Given these criteria, we were able to extract time series for the four-area dorsal stream model in 13 out of the 16 subjects and for all other models in 12 out of the 16 subjects (see Tables S2 and S3 for the coordinates of all regions in all participants). In the remaining four subjects, one of the areas in the model could not be defined due to the lack of an activation that met our criteria. These subjects therefore had to be excluded from the DCM analysis.

(ii) Definition of anatomical connections

Meta-analyses of primate connectivity data have shown that intra-hemispheric connections between visual areas are almost always reciprocal (Kötter & Stephan 2003). Also, macaque (Abel et al. 2000; Van Essen et al. 1982; Zeki 1970) and human (Clarke & Miklossy 1990; Van Essen et al. 1995; Van Valkenburg 1913) visual areas, with the notable exception of area V1, possess rich reciprocal callosal connections with their homotopic counterparts in the opposite hemisphere. Given these empirical facts, we assumed (i) reciprocal intra-hemispheric connections and (ii) reciprocal inter-hemispheric connections between homotopic areas in our model, except for the cuneus (CUN) in the dorsal stream model. The reason for omitting inter-hemispheric connections between left and right CUN was that an anatomical evaluation of our SPM results by means of a probabilistic cytoarchitectonic atlas (Eickhoff et al. 2005) indicated that activated CUN voxels were likely located in V1 (see Supplemental Material for details). However, the assumption of absent inter-hemispheric connections at the level of CUN was checked in subsequent models that did include these inter-hemispheric connections (Tables S7, S13) and was found to be appropriate.

(iii) Definition of driving inputs

We modeled the peripheral stimulus presentation by allowing all stimuli to directly induce activity in contralateral LG (ventral stream model) or contralateral CUN (dorsal stream model), respectively, regardless of task. These areas showed a main effect of visual field in the SPM analysis (see Supplemental Material for details). Since each stimulus only lasted for 150 ms, these inputs were represented as trains of events (delta functions).

(iv) Definition of modulatory inputs

The above choices of regions, connections and driving inputs resulted in the construction of basic models for the ventral and dorsal stream (Fig. 2A shows the basic model for the ventral stream). The final and critical step was to extend these models by modulatory inputs which change connection strengths as a function of the relevant experimental factors, i.e. task demands and visual field of stimulus presentation. Any given connection in the model could be influenced by four potential modulatory causes: The connection strength could depend (a) only on the visual field of stimulus presentation (the S model, for stimulus-dependent), (b) only on whether a specific task is performed or not (T model, for task-dependent), (c) on both the task and the visual field, but independently of each other (the T+S model), or (d) on both the task and the visual field, but in a conditional fashion, i.e. the connection strength is only modulated by task if the stimulus was presented in a particular visual field (the T×S model). These possibilities equally exist for inter- and the intra-hemispheric connections. We performed an exhaustive model comparison, systematically comparing all combinations of how inter- and intra-hemispheric connections could be changed as described above (Fig. 2B). Figs. 2C,D exemplify how combinations of the LD task and left/right visual fields yield 16 ventral stream models; analogous combinations, using SD as task, were chosen for the dorsal stream. However, we initially constrained the combinations by only allowing for modulation of the forward intra-hemispheric connections. This constraint was subsequently evaluated in additional models (Tables S7, S13) and found to be appropriate. Overall, this combinatorial approach resulted in 64 different DCMs per subject (16 DCMs each for the four-area and six-area models of both the ventral and dorsal stream). We generally refer to any specific models by first listing the modulation of the inter- and then that of the intra-hemispheric connections (compare Fig. 2B). For example, T/S is the model where inter-hemispheric connections are modulated by the task and the intra-hemispheric connections are modulated by the stimulus properties (i.e. visual field). All modulatory inputs were modeled as box-car inputs of 24 s duration.

Figure 2.

A. Basic structure of the four-area ventral stream model, comprising the reciprocally connected lingual gyrus (LG) and fusiform gyrus (FG) in both hemispheres. Due to the non-foveal stimulus presentation, stimuli in right (RVF) and left (LVF) visual field drive contralateral LG activity. During the instruction periods, bilateral visual field input was provided for 6 s; this was modeled as a box-car input affecting LG in both hemispheres (not shown here). The basic four-area model for the dorsal stream, comprising cuneus and superior parietal gyrus in both hemispheres, was constructed in an analogous fashion (see main text and Fig. 6).

B-D. Schema of how 16 variants of the ventral stream model (B) were constructed by systematically combining four different types of modulatory inputs for intra-hemispheric (C) and inter-hemispheric (D) connections. The strength of a connection could depend (a) only on the visual field of stimulus presentation (the S model, for stimulus-dependent), (b) only on whether a specific task is performed or not (T model, for task-dependent), (c) on both the task and the visual field, but independently of each other (the T+S model), or (d) on both the task and the visual field, but in a conditional fashion, i.e. the connection strength is only modulated by task if the stimulus was presented in a particular visual field (the T×S model). These possibilities equally exist for inter- and the intra-hemispheric connections. However, we initially constrained the combinations by only allowing for modulation of the forward intra-hemispheric connections. This constraint was subsequently evaluated in additional models and found to be appropriate (see main text). Overall, this combinatorial approach resulted in 64 different DCMs per subject (16 DCMs each for the four-area and six-area models of both the ventral and dorsal stream). Models are generally referred to by first listing the modulation of inter-hemispheric connections, followed by the modulation of intra-hemispheric connections (B). Analogous model variants were constructed for the dorsal stream, using SD as task.

DCM uses a gradient ascent procedure on the log posterior to compute MAP estimates (Friston et al. 2003). Given the relatively complex connectivity of our models, we took two steps to ensure convergence and validity of the parameter estimates. First, we used an extended version of the original DCM estimation scheme as described by Friston et al. (2003), using a Fisher scoring scheme with Levenburg-Marquardt regularization in the expectation maximization algorithm. This proved to ensure a robust behavior of the gradient ascent scheme. Second, in an associated methodological study (Stephan et al., in preparation), we performed a systematic series of simulations to test whether the structure of our models, in particular the multiple occurrence of identical modulatory inputs, could (i) influence the results from the BMS procedure and/or could (ii) lead to systematic bias in the sense that modulatory parameters of interest were overestimated. For the models presented here, we investigated this issue using simulations and found that even at high levels of noise (i) BMS reliably chose the correct model and that (ii) comparisons of modulations of inter-hemispheric connections were not biased.

Bayesian Model Selection (BMS)

Assessing model goodness is a central theme in statistics. Importantly, neither can one make statistical inferences about “absolute” model fit nor is model fit the only criterion to take into account when comparing models. As for the first point, all inferential statements about “model fit” or “variance accounted for” are relative to some null or reference model, e.g. R² values in regression models (Kleinbaum et al. 1988) or chi square tests in SEM (Penny et al. 2004b). Concerning the second point, model fit is a monotonic function of model complexity; overly complex models, however, will overfit the data and show inferior generalizability (Pitt & Myung 2002).

In this study, we used BMS to decide which DCM was optimal. BMS not only takes into account the relative fit of competing models but also their relative complexity (number of free parameters, functional form). It rests on the so-called “model evidence”, i.e. the probability p(y|m) of the data y given a particular model m (Raftery 1995). Usually, the model evidence cannot be determined analytically, therefore approximations are needed. For DCM, two suitable approximations are the Bayesian Information Criterion (BIC) and Akaike Information Criterion (AIC) (Penny et al. 2004a). These two approximations are both biased but in opposite ways: BIC tends to prefer simpler models whereas AIC tends to favor more complex models. This can lead to disagreement between the two approximations about which model should be favored. The general convention, to which we also adhere in this paper, is that, for any pairs of models m_i and m_j to be compared, a selection is only made if AIC and BIC concur. The decision is then based on that approximation which gives the more conservative Bayes factor (BF):

$${\mathit{BF}}_{ij}=\frac{p(y\mid m_i)}{p(y\mid m_j)}$$ (2)

An established convention is to prefer one model over another if the BF is larger than 3 (“positive evidence”, Raftery 1995).

When determining the optimal model for a group of individuals by BMS, it is likely that the optimal model will vary to some degree across subjects. Since model comparisons from different individuals are statistically independent, a group Bayes factor can be computed by multiplying the individual Bayes factors (where k is an index across subjects; see Stephan & Penny 2006):

$${\mathit{GBF}}_{\mathit{ij}}=\prod _k{\mathit{BF}}_{\mathit{ij}}^k$$ (3)

For each subject of our group, we first performed pairwise comparisons between all models and then computed the group Bayes factors (GBF) across subjects. However, GBFs can be misleading in the presence of strong outliers. Therefore, we additionally evaluated the number of comparisons for which the BF passed the threshold for positive evidence for either of the compared models. These numbers give a “positive evidence ratio” (PER) which serves as a complementary measure of which model is optimal at the group level (Stephan & Penny 2006).

Second-level analysis of model parameters

Once the best model was identified, the next step was to find which of the modeled processes were expressed consistently across subjects. We performed a classical second-level (between-subject) inference by applying a one-sample t-test to the corresponding MAP estimates from the individual DCMs. We performed these tests separately for each parameter of the fixed connections, modulatory changes of connections and driving inputs of the optimal model. We adopted a conservative procedure by using two-sided t-tests and a statistical threshold of p<0.05, with Bonferroni correction within each parameter class. This is analogous to the typical method for inference in conventional neuroimaging analyses, i.e. correcting for multiple comparisons within (but not across) several chosen statistical contrasts.

General methodological remarks

Before presenting the modeling results in detail, a few general methodological comments may be helpful. First, our models are agnostic to whether inter-hemispheric interactions are mediated through the corpus callosum or through other commissural tracts. However, this is neither critical for the definition of our models nor for the conclusions we draw from them. Second, our modeling results are not confounded by differences in task difficulty because the tasks were carefully matched for error rates (see Results). Third, in our factorial design, the experimental factors (visual field, task and response hand) were altered in a blocked fashion. Such a design differs from traditional behavioral paradigms in lateralization research that mostly use randomized visual field presentation in order to (i) minimize eye movements away from fixation and (ii) prevent attentional bias to contralateral visual space during prolonged stimulation of one hemisphere (Klein et al. 1976). In this study, we verified good fixation by online monitoring of eye movements using an infrared video system (see Stephan et al. 2003). Also, we controlled for potential effects of attentional bias by ensuring that the visual field of stimulation was perfectly balanced across all conditions of our factorial design. Generally, for studying lateralized processes with fMRI, blocking the visual field has three advantages: (i) for difficult tasks like ours, error rates can be kept at acceptable levels, (ii) the statistical power of the fMRI analysis is considerably higher, and (iii) whereas sustained covert spatial attention during a blocked design is not asymmetrically implemented in the brain (Tootell et al. 1998; Yantis et al. 2002), randomizing the visual field requires the subjects to engage continuously in attentional reorienting, a process that is known to induce activity lateralized to the right hemisphere (Corbetta & Shulman 2002). The last point highlights that blocked vs. randomized stimulation may lead to different mechanisms of stimulus processing; this should be kept in mind when comparing the results of the present study to previous studies which used randomized visual field stimulation.

Finally, please note that we explicitly focus on investigating IHI in separate ventral and dorsal streams models. The reason for this is that the conventional analysis based on the general linear model (GLM; see Fig. 1) showed a fairly clean separation between activations induced by letter judgments (confined to the ventral stream) and activations induced by spatial judgments (restricted to the dorsal stream). GLMs and DCMs are tightly linked in that they are both generative models of the same observed data, albeit differing in the explanations offered for how these observations were caused (compare Figs. 2&4 in Stephan 2004). Critically, the definition of any DCM depends on the results of a preceding GLM analysis that localizes those system elements (i.e. brain areas) where the experimental manipulations led to significant changes in BOLD signal. These local signal changes can subsequently be explained mechanistically by a DCM in terms of connections between the system elements and their modulation by experimental factors. In our case, the results of the GLM analysis (Fig. 1) suggested that it would be more appropriate to investigate IHI in separate ventral and dorsal streams models rather than constructing a large joint model. We verified this in an additional analysis in which we combined ventral and dorsal stream models into a joint model and tested, using BMS, whether one should include or exclude connections between the two streams. We found that treating ventral and dorsal stream models as separate networks was clearly more appropriate for the present paradigm (GBF > 10¹⁵; see Supplemental Material for details). We therefore only report the results from separate ventral and dorsal stream models in this article.

Figure 4.

Maximum a posteriori (MAP) parameter estimates (y-axis) for modulation of inter-hemispheric connections between left and right LG by letter decisions conditional on visual field (LD|VF) in all subjects (x-axis). Dark gray: modulation of left→right connection; light gray: modulation of right→left connection. The consistent asymmetry across subjects is obvious and highly significant (p=0.0007). Compare Fig. 3 and Table S5.

Results

Behavioral results

The tasks were designed to be of comparable difficulty as indexed by error rates. Indeed, analysis of variance of the behavioral responses during scanning demonstrated similar error rates between tasks (letter decisions: 8.5 ± 1.0 %; visuospatial decisions: 10.4 ± 2.0 %; p>0.196). However, the participants needed more time for letter decisions (686 ± 21 ms) than for visuospatial judgments (612 ± 28 ms) (p<0.001). For both error rates and reaction times, neither the main effects of visual hemifield or hand nor any of the interactions between the three factors were significant (see Stephan et al. 2003).

Construction of DCMs based on statistical parametric maps

In this study we re-analyzed the data by Stephan et al. (2003) using SPM2 to enable the application of DCM and BMS. To motivate the construction of our DCMs, we briefly summarize task-dependent activations in visual areas as obtained from the SPM analysis. All results described in this section are reported at p<0.05, whole-brain cluster-level corrected with a standard p<0.001 voxel-level cut-off.

Despite physically identical stimuli and matched task difficulty, a comparison of LD and SD tasks resulted in strongly lateralized activation patterns (Fig. 1, Table S1). Contrasting LD vs. SD, we found left-lateralized activations in several visual areas of the ventral stream, particularly in left fusiform gyrus (FG), left middle occipital gyrus (MOG) and bilateral lingual gyrus (LG) (Figs. 3, 5). This is in good accordance with other studies of letter processing (see Jobard et al. 2003 for review). In the opposite contrast, SD vs. LD, right-lateralized activations were found in dorsal stream areas, i.e. right superior parietal gyrus (SPG) and bilateral posterior angular gyrus (PAG; Fig. 6). An additional activation in right anterior parietal cortex (supramarginal and postcentral gyri; Fig. 1) substantially overlapped with somatosensory area 2 (according to a probabilistic cytoarchitectonic atlas, Eickhoff et al. 2005), and can thus not be regarded as a purely visual region.

Figure 3.

Summary of the group results for the optimal four-area ventral stream model (T×S/T). This model indicates that inter-hemispheric connections are modulated by the letter decision (LD) task, but conditional on the visual field of stimulation (LD|VF). Furthermore, there is a highly significant asymmetry in the strength of inter-hemispheric modulations (see Results). This indicates asymmetric information transfer from the non-dominant right to the dominant left hemisphere (red arrows). The average modulatory parameter estimates ± standard errors are shown alongside the modulatory inputs. Dark gray: significant effects surviving Bonferroni-correction; light gray: significant effects not surviving Bonferroni-correction; dotted lines: non-significant effects (see Table S5 for details). For clarity, this figure only shows the modulatory parameters; the values of all other parameters can be found in Table S6. The insets display the results from the random effects SPM analysis together with the group coordinates of the regions included in the model. Activations in left and right LG and left LG are shown at a threshold of p<0.05 cluster-level corrected (with p<0.001 voxel-level cut-off). Left and right LG activations are additionally masked by the main effect of visual field (RVF>LVF and LVF>RVF, respectively). The right FG activation is displayed at an uncorrected threshold of p<0.01.

Figure 5.

Summary of the group results for the optimal six-area ventral stream model (T×S/T). As in the four-area model (see Fig. 3), inter-hemispheric connections are modulated by the LD task, but conditional on the visual field of stimulation (LD|VF). This modulation is strongly asymmetric for LG and FG inter-hemispheric connections (see Results), indicating asymmetric information transfer from the non-dominant right to the dominant left hemisphere (red arrows). Notably, however, inter-hemispheric connections of MOG did not exhibit any significant modulation (compare Table S9). See legend to Fig. 3 for abbreviations and conventions.

Figure 6.

Summary of the group results for the optimal four-area dorsal stream model (T/T). Inter-hemispheric connection strengths were bidirectionally modulated by spatial decisions (SD), regardless of the visual field of stimulation. This task-dependent modulation symmetrically increased inter-hemispheric connections of SPG (Table S11). This connectivity pattern in the dorsal stream fits the predictions by the hemispheric recruitment theory. See legend to Fig. 3 for conventions. The insets display the results from the random effects SPM analysis together with the group coordinates of the regions included in the model (p<0.05 cluster-level corrected with p<0.001 voxel-level cut-off; CUN activations are additionally masked by the main effect of task, LD>SD).

Based on these results, we constructed a basic four-area ventral stream model for the LD task that comprised LG and FG in both hemispheres (Fig. 3A). The peripherally presented visual stimuli entered the system by directly affecting contralateral LG (note that LG showed a significant main effect of visual field; see insets in Fig. 3). The induced activity was then allowed to spread along reciprocal intra-hemispheric connections between LG and FG and reciprocal inter-hemispheric connections between left/right LG and left/right FG. From this basic model, 16 variants (Fig. 3B) were created by allowing that intra- (Fig. 3C) and inter-hemispheric connections could independently be modulated by one of four potential causes (Fig. 3D; see Methods section for more details on the construction of model variants). Specifically, the strength of a connection could depend (a) only on the visual field of stimulus presentation (the S model, for stimulus-dependent), (b) only on whether a specific task is performed or not (T model, for task-dependent), (c) on both the task and the visual field, but independently of each other (the T+S model), or (d) on both the task and the visual field, but in a conditional fashion, i.e. the connection strength is only modulated by task if the stimulus was presented in a particular visual field (the T×S model). We generally refer to any specific models by first listing the modulation of the inter- and then that of the intra-hemispheric connections (compare Fig. 3B). Another 16 variants of a six-area ventral stream model, which additionally included MOG bilaterally (Fig. 5), were constructed according to the same principles.

For the dorsal stream, an equivalent procedure was adopted although the basic connectivity layout was slightly different (see Methods for details). Here, we initially focused on connections between left and right SPG because this area has previously been implicated in IHI during visuospatial tasks (Iacoboni & Zaidel 2004). Since the SD vs. LD contrast did not activate any early visual area, we lacked an input site where stimuli would enter the system. We therefore complemented the dorsal stream model by a “neutral” input area, i.e. a subpart of the cuneus (CUN; see Fig. 6). Bilaterally, this area showed no significantly different activation between LD and SD tasks (p>0.05 uncorrected) but responded to visual stimuli in a hemifield-specific fashion (p<0.05 corrected). 16 variants of this basic four-area dorsal stream were constructed by allowing for context-dependent modulation of connection strengths, following the same principles used for construction of the ventral stream models (c.f. Fig. 2). An extended six-area dorsal stream model additionally included PAG in both hemispheres, giving rise to further 16 model variants. Please see Supplemental Material for details of how individual time series were extracted.

DCM: ventral stream

Comparing all 16 variants of the four-area ventral stream model across all subjects by BMS, the optimal model was found to be the T×S/T model (see Methods and Fig. 3B). According to this model, inter-hemispheric connections were modulated by the LD task but conditional on the visual field of stimulus presentation (LD|VF), whereas intra-hemispheric connections depended on the LD task alone. Table S4 shows the subject-specific Bayes factors for comparing the T×S/T model with the other 15 models and the resulting group Bayes factors. Across the group the evidence for the optimal T×S/T model was about 8 times higher as that of the second-best model (T/T×S), and the positive evidence ratio (PER) was 3:1. The results from the statistical group analysis, implemented as one-sample t-tests of the modulatory parameters from the optimal T×S/T model, are summarized by Fig. 3. Table S5 lists the estimates of the modulatory parameters for all subjects individually and the statistical results. Note that DCM parameters correspond to rate constants of the modeled neurophysiological processes and are thus given in Hz (see Methods).

The analysis of the optimal T×S/T model showed that right→left inter-hemispheric connections were significantly strengthened during LD whenever the stimulus was presented in the left visual field (LVF) and was thus initially received by the right hemisphere (Figs. 3, 4): The average rate constants for the modulation of the right→left inter-hemispheric connections by LD|LVF were 0.25 ± 0.04 Hz (right LG→left LG; p=0.0001) and 0.12 ± 0.02 Hz (right FG→left FG; p=0.0006), respectively. These effects were strong, corresponding to an increase in connectivity by a factor of 2.3 (LG) and 3.0 (FG) compared to the fixed (intrinsic) connection strengths (Table S6) and surviving Bonferroni correction for multiple comparisons. Notably, this modulation of inter-hemispheric connections was highly asymmetric, with modulatory effects being virtually absent for the left→right connections. Both at the level of LG and FG, this asymmetry was highly significant (p=0.0007 and p=0.0047; Table S5) and consistently expressed across subjects (Fig. 4). Altogether, these findings match exactly the predictions from the information transfer hypothesis (see Discussion): Stimulus information is transferred from the non-specialized to the specialized hemisphere, but only when the LD task is required and the stimulus information is presented to the non-specialized hemisphere.

Three things remain to be mentioned about the four-area model. Intra-hemispheric LG→FG connections were also strengthened, although by LD alone, and particularly in the left hemisphere (Fig. 3). Second, neither the fixed connection strengths nor the strengths of direct inputs showed any significant hemispheric asymmetries (Table S6). This emphasizes the notion that the ventral stream network is symmetric per se and that hemispheric differences are induced by the specific demands of the LD task. Finally, to test the basic assumptions underlying our models, we constructed and fitted additional variants of the basic structure of the T×S/T model for each subject. The first variant modeled additional changes in connection strengths by SD as well, i.e. inter-hemispheric connections were modulated by both LD|VF and SD|VF, and intra-hemispheric connections were modulated by both LD and SD. The second alternative model allowed for a modulation of the backward (FG→LG), instead of the forward, intra-hemispheric connections. A third alternative included a modulation of both forward (LG→FG) and backward (FG→LG) intra-hemispheric connections. BMS indicated that none of these alternative models came close to the T×S/T model: The group Bayes factors in favor of the T×S/T model were larger than 10¹², 10⁵⁸ and 10⁸, respectively (Table S7). Finally, an additional model comparison was performed to address a reviewer's concern whether inter-hemispheric connections between left and right LG exist in the human brain and should therefore be part of the model (see Discussion). Comparing the original T×S/T model against one without inter-hemispheric connections between left and right LG, we found the former to be superior (group Bayes factor > 10⁸). Altogether, these results confirmed that our initial choice of the basic model structure (Fig. 2) was sensible.

An extended six-area model of the ventral stream, which additionally included left and right MOG, gave very similar results (Fig. 5). The BMS procedure confirmed the same model, i.e. T×S/T, to be optimal. Although the group Bayes factor was marginally in favor of the similar T×S/T+S model (group BF = 0.798), this was due to a single outlier subject (see Table S8). In fact, in 10 out of 12 subjects the T×S/T was strongly preferred (with one subject showing no clear difference between the two models), resulting in a PER of 10:1. The statistical group analysis of the T×S/T model parameters replicated all results from the four-area model (compare Figs. 3, 5 and Tables S5, S9). Again, we found a strongly asymmetric modulation of inter-hemispheric connections: Right→left connections were significantly increased by LD|VF (LG: 0.20 ± 0.04 Hz, p=0.0003; FG: 0.07 ± 0.02 Hz, p=0.0012), whereas the corresponding modulation of left→right connections was significantly weaker (LG: p=0.0019; FG: p=0.0121). The other findings from the four-area model (stronger LD-modulation of the LG→FG connection in the left hemisphere, symmetric intrinsic connections and driving inputs) were also replicated. Importantly, no significant modulations were observed for inter-hemispheric connections between left and right MOG (right→left: 0.01 ± 0.01 Hz; p=0.1605; left→right: 0.00 ± 0.01 Hz, p=0.9725). Given this result, we focus on the four-area model as a sufficient representation of IHI in the ventral stream throughout the rest of this article.

DCM: dorsal stream

BMS among all 16 variants of the four-area dorsal stream model across all subjects indicated that the T/T model was optimal (Figs. 3B, 6; Table S10). According to this model, both intra-hemispheric connections and inter-hemispheric connections between left and right SPG were modulated by the SD task alone, regardless of the visual field of stimulation. Table S10 reports the subject-specific Bayes factors for comparing the T/T model with all other models. Across the group, the evidence for this model was about 10⁷ times higher as that for the second-best model (T/T×S), and the PER was 2:1. The statistical group analysis of the optimal model showed that inter-hemispheric SPG connections were significantly strengthened by the SD task in both directions (Fig. 6, Table S11), with average rate constants of 0.22 ± 0.06 Hz (right→left; p<0.0018) and 0.15 ± 0.06 Hz (left→right; p<0.0160). These modulations, corresponding to an increase in connectivity by a factor of 1.9 (right→left) and 2.9 (left→right), respectively, compared to the fixed (intrinsic) connection strengths alone (Table S12), were not significantly different from each other (p=0.2073). This connectivity pattern, i.e. symmetric enhancement of inter-hemispheric connections by task demands alone, matches the predictions from the hemispheric recruitment hypothesis (see Discussion): Regardless of which hemisphere initially receives the stimulus during the SD task, it increases its connectivity with the opposite hemisphere to recruit additional processing resources.

There are three more interesting aspects of the four-area dorsal stream model. Intra-hemispheric CUN→SPG connections were also strengthened by SD alone. This effect was significant in the right hemisphere (0.16 ± 0.04 Hz; p=0.0017), but not in the left hemisphere (0.11 ± 0.07 Hz). Second, as for the ventral stream model, neither the fixed connections nor the direct inputs showed any significant hemispheric asymmetries (see Table S12). Third, we tested the validity of the assumptions underlying the dorsal stream model by fitting five additional DCM variants for each subject and comparing these alternatives against the optimal T/T model. The first three alternatives were analogous to the alternative models tested for the ventral stream model (see above), i.e. (i) additional modulation of all connection strengths by LD, (ii) modulation of the backward (SPG→CUN), instead of the forward (CUN→SPG), intra-hemispheric connections, and (iii) modulation of both forward and backward intra-hemispheric connections. Two further alternative models asked whether (iv) an inclusion of inter-hemispheric connections between left and right CUN and (v) an additional modulation of these connections by SD would give a better model. The BMS procedure showed that all these alternative models were inferior to the T/T model (see Table S13). As with the ventral stream model, these additional comparisons corroborated our basic model structure.

The six-area dorsal stream model additionally included left and right PAG (Fig. 7). As with the four-area model, the BMS procedure found the T/T model to be optimal. The evidence for this model was about 2.8×10⁴ times higher as that of the second-best model (T/T×S; Table S14). The statistical group analysis of the T/T model parameter estimates replicated all results from the four-area model (Table S15). In particular, we again found a symmetric increase in SPG inter-hemispheric connections during SD (right→left: 0.09 ± 0.03 Hz, p=0.0063; left→right: 0.06 ± 0.02 Hz, p=0.0129; non-significant difference: p = 0.1746). However, no significant modulatory effects were found for the inter-hemispheric connections between left and right PAG (right→left: 0.00 ± 0.01 Hz; p=0.9009; left→right: 0.02 ± 0.01 Hz, p = 0.1745). Given this finding, the rest of this article focuses on the four-area model as a sufficient representation of IHI in the dorsal stream.

Figure 7.

Summary of the group results for the optimal six-area dorsal stream model (T/T). As in the four-area dorsal stream model (Fig. 6), inter-hemispheric connection strengths were bidirectionally modulated by spatial decisions, regardless of the visual field of stimulation. This task-dependent modulation symmetrically increased inter-hemispheric connections of SPG, but not of PAG. Compare Table S15 and see legend to Fig. 3 for conventions.

Discussion

A mechanistic understanding of IHI is important, not only for basic neuroscience, but also for clinical disorders related to hemispheric specialization, e.g. neglect, aphasia, or schizophrenia (Corbetta & Shulman 2002; Shkuro et al. 2000; Woodruff et al. 1997). IHI strongly depends on cognitive context, such as the task performed and the stimuli processed. Models of IHI are needed that account for this context-dependency, particularly during lateralized tasks (Hellige 1990; Liederman 1998). Testing such models experimentally, however, is challenging. Despite the importance of callosal lesions studies, their interpretation can be difficult, e.g. due to plastic reorganization of brain function. Coherence analyses based on EEG can be useful (Schack et al. 2003), but suffer from low anatomical resolution. Invasive recording studies (Engel et al. 1991; Kiper et al. 1999) cannot usually be performed in humans and only sample a few locations.

Models of effective connectivity in neuroimaging could overcome these problems. However, a challenge is that IHI models are usually complex due to the reciprocal nature of inter-hemispheric connections and the multiple pathways by which the hemispheres can interact. For example, structural equation modeling (SEM) is problematic in this situation because the parameters required often outnumber the observed covariances (McIntosh & Gonzalez-Lima 1994; Penny et al. 2004b). One can constrain reciprocal connections to be identical (Rowe et al. 2002) or constrain the fitting procedure (McIntosh et al. 1994), but neither is optimal. This may explain the lack of IHI studies that combine modeling and neuroimaging. An exception is McIntosh et al. (1994) who used two right-hemispheric tasks engaging ventral and dorsal visual streams, respectively. While the activation pattern was surprisingly symmetric for both tasks, SEM demonstrated asymmetries of inter-hemispheric connections, with right-to-left connections being positive and stronger during both tasks.

Currently, three major theories of IHI exist which predict different patterns of inter-hemispheric connectivity during lateralized tasks with lateralized inputs, as in our paradigm. The first theory is based on the notion of information transfer (Poffenberger 1912). It has been investigated in animals (Bures & Buresova 1960), by studies of split-brain patients (Gazzaniga 2000; Funnell et al. 2000; Corballis et al. 2003; Pollmann & Zaidel 1998), and in healthy volunteers (Brown & Jeeves 1993; Marzi et al. 1991). A broadly accepted proposal is that, given a particular lateralized task, information transfer should be asymmetrically enhanced from the non-specialized to the specialized hemisphere to ensure most efficient processing (Barnett & Kirk 2005; Endrass et al. 2002; Nowicka et al. 1996; but see Larson & Brown 1997). In terms of effective connectivity, this hypothesis predicts that connections towards the dominant hemisphere should be positive and significantly stronger than connections away from it. Furthermore, this task-dependent increase in connectivity should be particularly pronounced when stimulus information is initially only available to the non-dominant hemisphere, e.g. by presenting a visual stimulus in the periphery of the contralateral hemifield. In summary, the concept of information transfer predicts an asymmetric modulation of inter-hemispheric connections by task, but conditional on the visual field of stimulus presentation.

A second account of IHI concerns the functional balance between hemispheres. Kinsbourne (1970) proposed inter-hemispheric inhibition, mediated by mutual inhibition between homotopic brain regions, as a general principle of brain function. From this view, specialization of one hemispheric is equivalent to it being superior in suppressing the opposite hemisphere. While inter-hemispheric inhibition has been found mainly in the context of motor (Ferbert et al. 1992; Meyer et al. 1995) and visuospatial tasks (Fink et al. 2000; Hilgetag et al. 2001; Kapur 1996; Vuilleumier et al. 1996), Kinsbourne (1970) suggested that it might underlie all lateralized processes, including language. Recent studies also implicated it in non-spatial visual processing (Sack et al. 2005; Walsh et al. 1998). Although different variants of this theory use somewhat different concepts of “inhibition” (Chiarello & Maxfield 1996), most models predict that inter-hemispheric connection strengths should be negative in both directions (e.g. Hilgetag et al. 1999; Levitan & Reggia 2000). This does not necessarily mean, however, that two areas which affect each other by inter-hemispheric inhibition necessarily show decreased activity: regional activations can co-exist with inter-hemispheric inhibition if other, e.g. intra-hemispheric, inputs to the areas of interest are positive and dominant in magnitude.

A third major theory of IHI concerns hemispheric recruitment: when is it advantageous to restrict information processing to a single hemisphere or to distribute the computational load across both hemispheres? Extending pioneering work on limited hemispheric processing capacities (Hellige & Cox 1976; Liederman 1986), Banich and colleagues showed in behavioral experiments that hemispheric recruitment occurs as a function of computational complexity and attentional demands of the task performed (Banich 1998; Weissman & Banich 2000; Belger & Banich 1998; Passarotti et al. 2002). They conjectured that if the neural resources in the hemisphere receiving a stimulus are insufficient for optimal processing, the benefits of distributing processing load across both hemispheres should outweigh the costs of transcallosal information transfer. Given a sufficiently demanding task, hemispheric recruitment also occurs during lateralized tasks, even when the dominant hemisphere receives the stimulus (Belger & Banich 1998). Banich and colleagues predicted that hemispheric recruitment necessitates a tight coordination of processes in both hemispheres and that this should be reflected by a bidirectional and task-dependent increase in inter-hemispheric connectivity (Banich 1998; Weissman & Banich 2000).

Here, we have demonstrated how one can disambiguate among different candidate mechanisms of IHI for a given lateralized task, using DCM and BMS. Our modeling approach gave three main results. First, in the ventral stream, inter-hemispheric connections were asymmetrically modulated by letter processing but conditional on the stimulated VF (Figs. 3, 5): Right→left connections alone significantly increased during LD, but only with LVF stimulation. This is exactly the connectivity signature predicted by the information transfer hypothesis for a left-lateralized task. This asymmetric modulation of connectivity was observed for LG and FG, but not MOG. According to the model, MOG activation during LD is caused through its intra-hemispheric connections with LG and FG. We therefore conclude that MOG is not a key structure for IHI in letter processing. Second, in the dorsal stream, inter-hemispheric connections were bidirectionally modulated by spatial decisions, regardless of the stimulated VF (Figs. 6, 7). This is the connectivity signature predicted by the hemispheric recruitment hypothesis. As in the ventral stream model, the contextual modulation of inter-hemispheric connections was spatially specific: it was found for SPG, but not PAG. These results concur with the findings of Iacoboni & Zaidel (2004) who identified the right SPG as a key structure in IHI during a visuo-motor paradigm. Third, in all our models, neither the fixed (intrinsic) connection strengths nor the strengths of LVF/RVF inputs showed any asymmetries (Tables S6, S12). Instead, as described above, it is the specific task requirements that dynamically reconfigure connection strengths. This emphasizes the context-dependence of IHI and relates it to other cognitive processes in which connectivity changes mediate contextual effects (Büchel et al. 1999; Fries et al. 2001; Kastner & Ungerleider 2000; McIntosh 2000). It follows that the IHI mechanisms identified by the present study are probably not invariant properties of ventral and dorsal streams but are specific for the particular paradigm used. For other tasks, different mechanisms might be employed by the two streams (for example, see Hilgetag et al. 2001). The approach presented here introduces a general procedure, based on DCM and BMS, how the most likely mechanism of IHI during a specific cognitive task can be determined.

The structure of the optimal ventral and dorsal stream models differed with regard to the existence of inter-hemispheric connections at the level of the input areas. This was a deliberate choice because human post mortem tracing studies indicated that V1 (CUN in the dorsal stream model) lacks callosal connections while V2 (LG in the ventral stream model) is connected callosally (see Methods). Although these findings are not yet convincingly replicated by diffusion tensor imaging (DTI; Dougherty et al. 2005), this is likely due to the lower sensitivity of DTI as compared to anatomical studies. Nevertheless, we tested the validity of our assumptions about anatomical connectivity in additional models. Both removal of connections between left and right LG in the optimal ventral stream model and inclusion of connections between left and right CUN in the optimal dorsal stream model decreased model goodness (see Results for details). This emphasizes that the different mechanisms embodied by ventral and dorsal stream models are not due to invalid assumptions about the connectivity of the input-receiving areas, but are caused by the specific task requirements.

Altogether, our results provide direct neurophysiological evidence, in terms of effective connectivity, for asymmetric information transfer and hemispheric recruitment in ventral and dorsal streams, respectively, during a paradigm inducing task-driven lateralization. Furthermore, we have clarified the role of specific visual areas for IHI: LG and FG (but not MOG) are critical for IHI during letter processing, and SPG (but not PAG) during visuospatial processing. We hope that this study will be a useful starting point for future investigations of IHI in the human brain using dynamic system models of functional imaging data.